Model Size Reduction Research Articles

LRP‐based network pruning and policy distillation of robust and non‐robust DRL agents for embedded systems

SummaryReinforcement learning (RL) is an effective approach to developing control policies by maximizing the agent's reward. Deep reinforcement learning uses deep neural networks (DNNs) for function approximation in RL, and has achieved tremendous success in recent years. Large DNNs often incur significant memory size and computational overheads, which may impede their deployment into resource‐constrained embedded systems. For deployment of a trained RL agent on embedded systems, it is necessary to compress the policy network of the RL agent to improve its memory and computation efficiency. In this article, we perform model compression of the policy network of an RL agent by leveraging the relevance scores computed by layer‐wise relevance propagation (LRP), a technique for Explainable AI (XAI), to rank and prune the convolutional filters in the policy network, combined with fine‐tuning with policy distillation. Performance evaluation based on several Atari games indicates that our proposed approach is effective in reducing model size and inference time of RL agents. We also consider robust RL agents trained with RADIAL‐RL versus standard RL agents, and show that a robust RL agent can achieve better performance (higher average reward) after pruning than a standard RL agent for different attack strengths and pruning rates.

An Auto-Encoder Based TinyML Approach for Real-Time Anomaly Detection

<div class="section abstract"><div class="htmlview paragraph">Condition monitoring plays a crucial role in the automotive space because they reduce the downtime and maintenance costs by preventing sudden catastrophic breakdown of vehicles. Condition monitoring solutions primarily monitor onboard sensor data to detect anomalies. However, with vehicles becoming more complex each day, the number of sensors to be monitored is also growing. This increases the data volume to be processed to make decisions. It is not feasible to send all the embedded sensor data captured to the cloud for processing. The network bandwidth, latency for transferring the data and finally the cost associated with data transfer are the factors which impede us from taking this approach. Much of the previous work done to address these issues has proposed Deep Learning driven onboard anomaly detection approaches. The deployment of these implementations requires power hungry and costly hardware with high computational resources. The recent advances in the field of Tiny Machine Learning have made it possible to design and deploy Deep Neural Networks on resource constrained low-cost embedded hardware. In this paper, we propose an Auto-Encoder based approach for anomaly detection in time-series vibration sensor data. The designed Auto-Encoder model has a 7.5 KB footprint and was finally deployed on a highly resource constrained ARM Cortex-M4 microcontroller with 256KB of SRAM and 1MB Flash. The model has been validated on the previous machine data. It has achieved an accuracy and precision close to 80%. Currently, by using post training quantization we are trading-off model accuracy for a reduction in model size. In future, we plan to use Quantization Aware Training which will help us in achieving even higher model accuracy.</div></div>

Effective multi-crop disease detection using pruned complete concatenated deep learning model

A significant threat to agriculture yield is crop disease. It leads to enormous losses for farmers and also has an impact economically. Leaves affected by certain diseases will exhibit unique characteristics, which will be utilized by deep learning frameworks to identify the diseases. In this paper, the Complete Concatenated Deep Learning (CCDL) architecture, a multi-crop disease detection model is proposed that is capable to classify the crop diseases irrespective of crops. In this architecture, Complete Concatenated Block (CCB) is introduced as a core functional unit. In this unit, the point convolution layer is positioned before every convolution layer to confine the number of parameters generated in the model. A complete concatenation path is invoked upon the convolution layers, contained within the CCB. It enhances the utilization of feature maps and helps to achieve better classification accuracy. The proposed architecture is trained using the reorganized Plant Village dataset. Later the trained model has been pruned for model size reduction, called Pruned Complete Concatenated Deep Learning model (PCCDL). This proposed architecture is delivered as three variants, in which the model PCCDL with Partial Standard Convolution Technique (PCCDL-PSCT) outperformed and achieved a higher classification accuracy of 98.14 % with a lesser model size of ∼10 MB.

Online Distance Field Priors for Gaussian Process Implicit Surfaces

Gaussian process (GP) implicit surface models provide environment and object representations which elegantly address noise and uncertainty while remaining sufficiently flexible to capture complex geometry. However, GP models quickly become intractable as the size of the observation set grows—a trait which is difficult to reconcile with the rate at which modern range sensors produce data. Furthermore, naïve applications of GPs to implicit surface models allocate model resources uniformly, thus using precious resources to capture simple geometry. In contrast to prior work addressing these challenges though model sparsification, spatial partitioning, or ad-hoc filtering, we propose introducing model bias online through the GP’s mean function. We achieve more accurate distance fields using smaller models by creating a distance field prior from features which are easy to extract and have analytic distance fields. In particular, we demonstrate this approach using linear features. We show the proposed distance field halves model size in a 2D mapping task using data from a SICK S300 sensor. When applied to a single 3D scene from the TUM RGB-D SLAM dataset, we achieve a fivefold reduction in model size. Our proposed prior results in more accurate GP implicit surfaces, while allowing existing models to function in larger environments or with larger spatial partitions due to reduced model size.

An Efficient Implementation of Convolutional Neural Network With CLIP-Q Quantization on FPGA

Convolutional neural networks (CNNs) have achieved tremendous success in the computer vision domain recently. The pursue for better model accuracy drives the model size and the storage requirements of CNNs as well as the computational complexity. Therefore, Compression Learning by InParallel Pruning-Quantization (CLIP-Q) was proposed to reduce a vast amount of weight storage requirements by using a few quantized segments to represent all weights in a CNN layer. Among various quantization strategies, CLIP-Q is suitable for hardware accelerators because it reduces model size significantly while maintaining the full-precision model accuracy. However, the current CLIP-Q approach did not consider the hardware characteristics and it is not straightforward when mapped to a CNN hardware accelerator. In this work, we propose a software-hardware codesign platform that includes a modified version of CLIP-Q algorithm and a hardware accelerator, which consists of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$5\times 5$ </tex-math></inline-formula> reconfigurable convolutional arrays with input and output channel parallelization. Additionally, the proposed CNN accelerator maintains the same accuracy of a full-precision CNN in Cifar-10 and Cifar-100 datasets.

CAP’NN: A Class-aware Framework for Personalized Neural Network Inference

We propose a framework for Class-aware Personalized Neural Network Inference (CAP’NN), which prunes an already-trained neural network model based on the preferences of individual users. Specifically, by adapting to the subset of output classes that each user is expected to encounter, CAP’NN is able to prune not only ineffectual neurons but also miseffectual neurons that confuse classification, without the need to retrain the network. CAP’NN also exploits the similarities among pruning requests from different users to minimize the timing overheads of pruning the network. To achieve this, we propose a clustering algorithm that groups similar classes in the network based on the firing rates of neurons for each class and then implement a lightweight cache architecture to store and reuse information from previously pruned networks. In our experiments with VGG-16, AlexNet, and ResNet-152 networks, CAP’NN achieves, on average, up to 47% model size reduction while actually improving the top-1(5) classification accuracy by up to 3.9%(3.4%) when the user only encounters a subset of the trained classes in these networks.

RDPNet: a single-path lightweight CNN with re-parameterization for CPU-type edge devices

Deep convolutional neural networks have produced excellent results when utilized for image classification tasks, and they are being applied in a growing number of contexts. Model inference on edge devices is challenging due to the unending complicated structures needed to improve performance, which adds a significant computing burden.According to recent research, the often utilized residual structure in models does not support model inference. The idea of structural reparameterization is put out to address this shortcoming. The RepVGG produced with this method is a high-performance, quick-inference single-path network. Even after reparameterization, the model still needs GPUs and other specialized computing libraries to accelerate inference, however this still has a limit on how quickly the model can infer at the edge. We construct RDPNet using depthwise separable convolution and structural reparameterization to further reduce model size and accelerate inference. When utilizing an Intel CPU, this is a straightforward network that may be utilized for inference. For re-parameterization, we specifically adopt Depthwise separable convolution as the basic convolution form. Create a multi-branch model for training on the training side, and then simplify it into a single-branch model that the edge devices can easily infer. Research demonstrates that compared to alternative lightweight networks that can attain SOTA performance, RDPNet offers a superior trade-off between accuracy and latency.

STAR-Lite: A light-weight scalable self-taught learning framework for older adults’ activity recognition

Recognizing activities for older adults is challenging as we observe a variety of activity patterns caused due to aging (e.g., limited dexterity, limb control, slower response time) or/and underlying health conditions (e.g., dementia). However, existing literature with deep learning methods has successfully recognized activities when the dataset contains high-quality annotations and is captured in a controlled environment. On the contrary, data captured in a real-world environment, especially with older adults exhibiting memory-related symptoms, varying psychological and mental health status, reliance on caregivers to perform daily activities, and unavailability of domain-specific annotators, makes obtaining quality data with annotations challenging; leaving us with limited labeled data and abundant unlabeled data. In this paper, we hypothesize that projecting the labeled data representations comprising a specific set of activities onto a new representation space characterized by the unlabeled data comprising activities beyond the limited activities in the labeled dataset would help us rely less on the annotated data to improve activity detection performance. Motivated by this, we propose STAR-Lite, a self-taught learning framework that involves a pre-training framework to prepare the new representation space considering activities beyond the initial labels in the labeled dataset. STAR-Lite projects the labeled data representations on the new representation space characterized by unlabeled data labels and learns higher-level representations of the labeled dataset while optimizing inter- and intra- class distances without explicitly using a computation hungry similarity-based approach. We demonstrate that our proposed approach, STAR-Lite (a) improves activity recognition performance in a supervised setting and (b) is feasible for real-world deployment. To enhance the feasibility of deploying STAR-Lite on devices with limited memory resources, we explore model compression techniques such as pruning and quantization and propose a novel layer-wise pruning-rate optimization technique that effectively compresses the network while preserving the model performance. The evaluation was performed using the Alzheimer’s Activity Recognition dataset (AAR) captured from 25 individuals living in a retirement community center with IRB approval (#Y18NR12035) using an in-house SenseBox infrastructure while concurrently assessing the clinical evaluation of the participants for dementia, and independent living. Our extensive evaluation reveals that STAR-Lite can detect activities with an F1-score of 85.12% despite 62% reduction in model size and 5% improvement of execution time on a resource constrained device.

Rapid Target Detection of Fruit Trees Using UAV Imaging and Improved Light YOLOv4 Algorithm

The detection and counting of fruit tree canopies are important for orchard management, yield estimation, and phenotypic analysis. Previous research has shown that most fruit tree canopy detection methods are based on the use of traditional computer vision algorithms or machine learning methods to extract shallow features such as color and contour, with good results. However, due to the lack of robustness of these features, most methods are hardly adequate for the recognition and counting of fruit tree canopies in natural scenes. Other studies have shown that deep learning methods can be used to perform canopy detection. However, the adhesion and occlusion of fruit tree canopies, as well as background noise, limit the accuracy of detection. Therefore, to improve the accuracy of fruit tree canopy recognition and counting in real-world scenarios, an improved YOLOv4 (you only look once v4) is proposed, using a dataset produced from fruit tree canopy UAV imagery, combined with the Mobilenetv3 network, which can lighten the model and increase the detection speed, combined with the CBAM (convolutional block attention module), which can increase the feature extraction capability of the network, and combined with ASFF (adaptively spatial feature fusion), which enhances the multi-scale feature fusion capability of the network. In addition, the K-means algorithm and linear scale scaling are used to optimize the generation of pre-selected boxes, and the learning strategy of cosine annealing is combined to train the model, thus accelerating the training speed of the model and improving the detection accuracy. The results show that the improved YOLOv4 model can effectively overcome the noise in an orchard environment and achieve fast and accurate recognition and counting of fruit tree crowns while lightweight the model. The mAP reached 98.21%, FPS reached 96.25 and F1-score reached 93.60% for canopy detection, with a significant reduction in model size; the average overall accuracy (AOA) reached 96.73% for counting. In conclusion, the YOLOv4-Mobilenetv3-CBAM-ASFF-P model meets the practical requirements of orchard fruit tree canopy detection and counting in this study, providing optional technical support for the digitalization, refinement, and smart development of smart orchards.

A High-Precision Water Body Extraction Method Based on Improved Lightweight U-Net

The rapid, accurate extraction of water body information is critical for water resource management and disaster assessment. Its data foundation was mostly provided by remote sensing images through deep learning methods. However, the methods still require the improvement of recognition accuracy and reduction of model size. As a solution, this paper proposed a new high-precision convolutional neural network for water body extraction. This network’s structural design is based on the assumption that the extraction effect of a convolutional neural network is independent from its parameters number, thus the recognition effect could be effectively improved through reasonable adjustment of the network structure according to characteristics of water bodies on high-resolution remote sensing images. It brings two critical improvements. Firstly, the number of downsampling layers was reduced to adapt to the low resolution of remote sensing imagery. Secondly, the bottleneck structure has also been updated to fit the decoder–encoder framework. The improved bottleneck structures were nested to ensure the transmission of water characteristics information in the model. In comparison with the other five commonly used networks, the new network has achieved the best results (average overall accuracy: 98.31%, parameter benefit value: 0.2625), indicating the extremely high practical value of this approach.

Cross-domain knowledge distillation for text classification

Most text classification methods achieve great success based on the large-scale annotated data and the pre-trained language models. However, the labeled data is insufficient in practice, and the pre-trained language models are difficult to deploy due to their high computing resources and slow inference speed. In this paper, we propose cross-domain knowledge distillation, where the teacher and student tasks belong to different domains. It not only acquires knowledge from multiple teachers but also accelerates inference and reduces model size. Specifically, we train the pre-trained language models on factual knowledge obtained by aligned Wikipedia text to Wikidata triplets and fine-tune it as the teacher model. Then we use the heterogeneous multi-teacher knowledge distillation to transfer knowledge from the multiple teacher models to the student model. Multi-teacher knowledge vote can distill knowledge related to the target domain. Moreover, we also introduce the teacher assistant to help distill large pre-trained language models. Finally, we reduce the difference between the source domain and target domain by multi-source domain adaptation to solve the domain shift problem. Experiments on the multiple public datasets demonstrate that our method can achieve competitive performance while having fewer parameters and less inference time.

Generic medical concept embedding and time decay for diverse patient outcome prediction tasks

SummaryMany fields, including Natural Language Processing (NLP), have recently witnessed the benefit of pre-training with large generic datasets to improve the accuracy of prediction tasks. However, there exist key differences between the longitudinal healthcare data (e.g., claims) and NLP tasks, which make the direct application of NLP pre-training methods to healthcare data inappropriate. In this article, we developed a pre-training scheme for longitudinal healthcare data that leverages the pairing of medical history and a future event. We then conducted systematic evaluations of various methods on ten patient-level prediction tasks encompassing adverse events, misdiagnosis, disease risks, and readmission. In addition to substantially reducing model size, our results show that a universal medical concept embedding pretrained with generic big data as well as carefully designed time decay modeling improves the accuracy of different downstream prediction tasks.

FastStereoNet: A Fast Neural Architecture Search for Improving the Inference of Disparity Estimation on Resource-Limited Platforms

Convolutional neural networks (CNNs) provide the best accuracy for disparity estimation. However, CNNs are computationally expensive, making them unfavorable for resource-limited devices with real-time constraints. Recent advances in neural architectures search (NAS) promise opportunities in automated optimization for disparity estimation. However, the main challenge of the NAS methods is the significant amount of computing time to explore a vast search space [e.g., <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1.6\times 10^{29}$ </tex-math></inline-formula> ] and costly training candidates. To reduce the NAS computational demand, many proxy-based NAS methods have been proposed. Despite their success, most of them are designed for comparatively small-scale learning tasks. In this article, we propose a fast NAS method, called FastStereoNet, to enable resource-aware NAS within an intractably large search space. FastStereoNet automatically searches for hardware-friendly CNN architectures based on late acceptance hill climbing (LAHC), followed by simulated annealing (SA). FastStereoNet also employs a fine-tuning with a transferred weights mechanism to improve the convergence of the search process. The collection of these ideas provides competitive results in terms of search time and strikes a balance between accuracy and efficiency. Compared to the state of the art, FastStereoNet provides <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$5.25\times $ </tex-math></inline-formula> reduction in search time and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$44.4\times $ </tex-math></inline-formula> reduction in model size. These benefits are attained while yielding a comparable accuracy that enables seamless deployment of disparity estimation on resource-limited devices. Finally, FastStereoNet significantly improves the perception quality of disparity estimation deployed on field-programmable gate array and Intel Neural Compute Stick 2 accelerator in a significantly less onerous manner.

A cell phone app for facial acne severity assessment.

Acne vulgaris, the most common skin disease, can cause substantial economic and psychological impacts to the people it affects, and its accurate grading plays a crucial role in the treatment of patients. In this paper, we firstly proposed an acne grading criterion that considers lesion classifications and a metric for producing accurate severity ratings. Due to similar appearance of acne lesions with comparable severities and difficult-to-count lesions, severity assessment is a challenging task. We cropped facial skin images of several lesion patches and then addressed the acne lesion with a lightweight acne regular network (Acne-RegNet). Acne-RegNet was built by using a median filter and histogram equalization to improve image quality, a channel attention mechanism to boost the representational power of network, a region-based focal loss to handle classification imbalances and a model pruning and feature-based knowledge distillation to reduce model size. After the application of Acne-RegNet, the severity score is calculated, and the acne grading is further optimized by the metadata of the patients. The entire acne assessment procedure was deployed to a mobile device, and a phone app was designed. Compared with state-of-the-art lightweight models, the proposed Acne-RegNet significantly improves the accuracy of lesion classifications. The acne app demonstrated promising results in severity assessments (accuracy: 94.56%) and showed a dermatologist-level diagnosis on the internal clinical dataset.The proposed acne app could be a useful adjunct to assess acne severity in clinical practice and it enables anyone with a smartphone to immediately assess acne, anywhere and anytime.

A Pruned Deep Learning Approach for Classification of Motor Imagery Electroencephalography Signals.

The Deep Learning (DL) approach has been gaining much popularity in recent years in the development of electroencephalogram (EEG) based Motor Imagery (MI) Brain-Computer Interface (BCI) systems, aiming to improve the performance of existing stroke rehabilitation strategies. A complex deep neural network structure has lots of neurons with thousands of parameters to optimize, and a great deal of data is often required to train the network and the training process can take an extremely long time. High training costs and high model complexity not only have negative impacts on the performance of the BCI system but also on its applicability to meet the real-time requirement to support the rehabilitation exercises of patients. To tackle the challenge, a contribution-based neuron selection method is proposed in this paper. A Convolutional Neural Network (CNN) based motor imagery classification framework is implemented, and a neuron pruning approach is developed and applied. The temporal and spatial features of EEG signals are captured by the CNN layers, and then the fast recursive algorithm (FRA) is applied to prune redundant parameters in the fully connected layers which reduces the computation cost of the CNN model without affecting its performance. The experimental results show that the proposed method can achieve up to 50% model size reduction and 67.09% computation savings.

LTH-ECG: Lottery Ticket Hypothesis-based Deep Learning Model Compression for Atrial Fibrillation Detection from Single Lead ECG On Wearable and Implantable Devices.

Atrial Fibrillation (AF) is a kind of arrhythmia, which is a major morbidity factor, and AF can lead to stroke, heart failure and other cardiovascular complications. Electrocardiogram (ECG) is the basic marker to test the condition of heart and it can effectively detect AF condition. Single lead ECG has the practical advantage for being small form factor and it is easy to deploy. With the sophistication of the current deep learning (DL) models, researchers have been able to construct cardiologist-level models to detect different arrhythmias including AF condition detection from single lead short-time ECG signals. However, such models are computationally expensive and require huge memory size for deployment (more than 100 MB to deploy state-of-the-art 34-layer convolutional neural network-based ECG classification model). Such models need to be significantly trimmed with insignificantly loss of its classification performance for deployment in practical applications like single lead ECG classification in wearable and implantable devices. We have found that classical deep learning model compression techniques like pruning, quantization are not capable of substantial model size reduction without compromising on the model performance. In this paper, we propose LTH-ECG, which is our novel goal-driven winning lottery ticket discovery method, where lottery ticket hypothesis (LTH)-based iterative model pruning is used with the aim of over-pruning avoidance. LTH-ECG reduces the model size by 142x times with insignificant loss of classification performance (less than 1 % test F1-score penalty). Clinical Relevance- LTH-ECG will enable practical deployment for remote screening of AF condition using single lead short-time ECG recordings such that patients can on-demand monitor AF condition remotely through wearable ECG sensing devices and report cardiological abnormality to the concerned physician. LTH-ECG acts as an early warning system for effective AF condition screening.

Deeply Tensor Compressed Transformers for End-to-End Object Detection

DEtection TRansformer (DETR) is a recently proposed method that streamlines the detection pipeline and achieves competitive results against two-stage detectors such as Faster-RCNN. The DETR models get rid of complex anchor generation and post-processing procedures thereby making the detection pipeline more intuitive. However, the numerous redundant parameters in transformers make the DETR models computation and storage intensive, which seriously hinder them to be deployed on the resources-constrained devices. In this paper, to obtain a compact end-to-end detection framework, we propose to deeply compress the transformers with low-rank tensor decomposition. The basic idea of the tensor-based compression is to represent the large-scale weight matrix in one network layer with a chain of low-order matrices. Furthermore, we propose a gated multi-head attention (GMHA) module to mitigate the accuracy drop of the tensor-compressed DETR models. In GMHA, each attention head has an independent gate to determine the passed attention value. The redundant attention information can be suppressed by adopting the normalized gates. Lastly, to obtain fully compressed DETR models, a low-bitwidth quantization technique is introduced for further reducing the model storage size. Based on the proposed methods, we can achieve significant parameter and model size reduction while maintaining high detection performance. We conduct extensive experiments on the COCO dataset to validate the effectiveness of our tensor-compressed (tensorized) DETR models. The experimental results show that we can attain 3.7 times full model compression with 482 times feed forward network (FFN) parameter reduction and only 0.6 points accuracy drop.

Hybrid Neural Networks for On-Device Directional Hearing

On-device directional hearing requires audio source separation from a given direction while achieving stringent human-imperceptible latency requirements. While neural nets can achieve significantly better performance than traditional beamformers, all existing models fall short of supporting low-latency causal inference on computationally-constrained wearables. We present DeepBeam, a hybrid model that combines traditional beamformers with a custom lightweight neural net. The former reduces the computational burden of the latter and also improves its generalizability, while the latter is designed to further reduce the memory and computational overhead to enable real-time and low-latency operations. Our evaluation shows comparable performance to state-of-the-art causal inference models on synthetic data while achieving a 5x reduction of model size, 4x reduction of computation per second, 5x reduction in processing time and generalizing better to real hardware data. Further, our real-time hybrid model runs in 8 ms on mobile CPUs designed for low-power wearable devices and achieves an end-to-end latency of 17.5 ms.

AFSense-ECG: Atrial Fibrillation Condition Sensing From Single Lead Electrocardiogram (ECG) Signals

In this paper, we propose AFSense-ECG, an intelligence-embedded single lead ECG sensor that is enabled with the ability of accurate detection of Atrial Fibrillation (AF) condition, which is the most common sustained cardiac arrhythmia and increased risk of stroke is higher with sub-clinical AF patients. AFSense-ECG acts like an early-warning sensor for AF condition detection. A processing unit (e.g., ESP32WROVERE microcontroller) integrated with off-the-shelf single lead ECG sensor like Alivecor or AD8232 embeds intelligence to the sensing system to augment for inferential sensing for empowering automated decision-making. AFSense-ECG captures the quasi-periodic nature of typical ECG signals with repetitive P-wave, QRS complex and T-wave patterns into its feature extraction and the representation learning process of model construction and learning rate optimization. Our empirical study validates the superiority of proposed ECG signal characteristics-based hyperparameter tuned ECG classification model construction. AFSense-ECG demonstrates F1-measure of 86.13%, where the current state-of-the-art methods report F1-measures of 83.70%, 83.10%, 82.90%, 82.60%, 82.50%, 81.00% over publicly available single lead ECG datasets of Physionet 2017 Challenge. Further, the proposed learning model for the inferential sensing is lean (approximately 25 times simpler in terms of total number of trainable parameters with reduced model size than relevant state-of-the-art model, where the state-of-the-art method with 83.70% F1-measure consists of 10474607 trainable parameters, and our proposed model consists of 433675 trainable parameters) and more effective (better F1-measure than the state-of-the-art methods), which enables us to construct affordable intelligent sensing system.

Compact N-gram Language Models for Armenian

Applications such as speech recognition and machine translation use language models to select the most likely translation among many hypotheses. For on-device applications, inference time and model size are just as important as performance. In this work, we explored the fastest family of language models: the N-gram models for the Armenian language. In addition, we researched the impact of pruning and quantization methods on model size reduction. Finally, we used Bye Pair Encoding to build a subword language model. As a result, we obtained a compact (100 MB) subword language model trained on massive Armenian corpora.